On September 11, 2026, Agibot (Zhiyuan Robotics) open-sourced GE-Act 2.0, a native world–action model pre-trained from scratch on robot manipulation data. Scaling joint training data from 300 hours to 30,000 hours raised the humanoid G1-OP's zero-shot success rate from 17.1% to 44.1%, and tasks achieving non-zero scores grew from 39 to 76 out of 100. The team notes: 300 hours is about 12.5 days; 30,000 hours is 3.4 years. Such scaling curves had been repeatedly validated for language models, but robotics lacked systematic evidence — until now.
World–Action Model vs. VLA
Standard vision-language-action (VLA) models map camera frames and language instructions directly to motor commands — treating physical interaction as text continuation — without explicitly modeling what happens when you touch the world. A world–action model (WAM) fills that gap: first predict how the scene should evolve, then decide how to act. But conventional video generation models are too slow and memory-hungry for real-time control.
GE-Act 2.0 splits the job into three stages:
1. CoAE — a control-first visual tokenizer compressing a 256×384 frame into 24 visual tokens (1/16 of DINOv3's representation), discarding background detail while preserving geometry, boundaries, and motion. On a controlled test of 4,000 manipulation samples, instruction-frame matching accuracy hit 97.95%, the best of five comparison representations. 2. SVP — a one-step visual planner that generates the imagined future in a single forward pass (no iterative denoising): near frames for immediate control, far frames for staying on course. Reported latency: ~104 ms per action sequence on an RTX 5090. 3. IDM — an inverse dynamics model that converts the imagined future into smooth joint commands.
The Validity Gap and KASO
Physical manipulation is inherently multi-solution: approaching a cup from the left or right are both correct, but demonstrations record only one path. If the planner imagines the right-side approach while demonstrations show the left, conventional training forces conflicting supervision, causing policies to freeze or average into useless motion — the validity gap.
KASO (Knowledge-Aligned Selective Optimization) addresses this: during training the planner generates multiple candidate futures, and only the one aligned with the real demonstration is used to update weights. Real-robot ablations show grasp success rising from 22.5% to 37.5% with KASO.
Four Data Scales, Three Lines of Evidence
The same recipe was used across 300 / 1,200 / 5,000 / 30,000 hours:
| Training scale | G1-OP success | G2-90D success | Non-zero tasks (G1-OP) | |---|---|---|---| | 300 h | 17.1% | 13.4% | 39 / 100 | | 1,200 h | 22.6% | 21.0% | 51 / 100 | | 5,000 h | 27.3% | 23.5% | 62 / 100 | | 30,000 h | 44.1% | 31.1% | 76 / 100 |
- Skills light up progressively: fine tasks like towel folding, cup nesting, and pen-cap insertion only show non-zero success at 30k hours.
- No saturation yet: the curve still climbs from 5k to 30k hours.
- Rich embodiments lift scarce ones: the wheeled G2-90D has under 2% of training data yet gained 17.7 percentage points, borrowing spatial dynamics learned from other robots.
- Mid-motion instruction change: while reaching for a green cup, the instruction switched to "take the blue cup" centimeters from contact. The arm paused ~1.5 s, left the original trajectory, and switched targets. Another test swapped hands mid-reach.
- Counter-intuitive binding: asked to "put the cup in the shoebox" (with a shoe and shoebox on the table), the model followed the literal instruction rather than the shoe-into-box visual association.
A data audit shows zero-shot success correlates log-linearly with per-skill training hours: table wiping (824 h) reaches 76.7% reliability; broom sweeping (64 h) stalls at 3.3%.
Why Not Adapt a Text-to-Video Model?
Most peers fine-tune an off-the-shelf text-to-video model with an action head, inheriting world priors from millions of hours of internet video. Agibot's stated objection is misaligned objectives: video models optimize for plausible-looking frames; robots need executable, precise actions. A two-centimeter finger error is invisible to the eye but decisive for grasping. Pre-training from scratch means collecting your own data — hence the three-layer architecture:
1. 39,000 h instruction–video for world-model pre-training (including 3,000 h embodiment-free video, no action labels). 2. 32,000 h robot trajectories for IDM pre-training (including 2,000 h of failed operations and deployment-returned data — no success labels required). 3. 30,000 h instruction–video–action for joint training.
The IDM is the pivot: given before/after frames, it recovers action supervision from unlabeled, instruction-free, even failed trajectories — deployment failures aren't wasted.
Qualitative Highlights
Open-Source Context: Unitree the Day Before
On September 10, Unitree open-sourced UnifoLM-WLA-1.0 (6B parameters), a single model covering tabletop and whole-body mobile manipulation with cross-task, cross-end-effector generalization. Agibot's GE-Act 2.0 followed on September 11. Evaluation was kept strict: no fine-tuning on test tasks, no demonstrations, pure real-robot zero-shot, across 100 atomic tasks, 20 skill categories, and two embodiments with unseen scenes, backgrounds, lighting, and object instances.
Industrial Context
Agibot's Spirit G2 now runs routinely on Longqi Technology's Nanchang factory line — described as the first scaled embodied-AI 3C precision-manufacturing deployment — with plans to expand to 100 units in Q3 2026. A Serbian plant with Minth Group started production on August 29, targeting 1,000–2,000 humanoids per year.
Unitree founder Wang Xingxing has defined embodied AI's "ChatGPT moment" as completing 80% of tasks via natural instructions in 80% of unseen scenarios, estimating 2–3 years at best, 5–10 at worst. Against that bar, 44.1% still falls more than half short — and production-line returned data is exactly the fuel for closing that gap.
What 44.1% Means
Agibot itself hedges: the experiments only demonstrate that, within its architecture, data system, and evaluation scope, more data improves performance on unfamiliar tasks. Untested: more embodiments, long-horizon tasks, and unstructured environments.
Three things to watch:
1. Whether the unsaturated 5k→30k curve continues at the next data scale. 2. Whether the under-2%-data, +17.7-point transfer replicates to third and fourth embodiments. 3. As these open models spread, software moats for small/medium robot makers shrink while data moats grow — strengthening upstream data collection and simulation.
References
1. Humanoids Daily, "AGIBOT Open-Sources GE-Act 2.0, Challenging Unitree in the Open Embodied AI Race" (2026-09-11) 2. Shenzhen Economic Daily / Duchuang, "Agibot releases GE-Act 2.0: 100x training data, higher zero-shot robot success" (2026-09-09) 3. Agibot official release and GE-Act 2.0 technical paper (arXiv:2609.05588) 4. Unitree UnifoLM-WLA-1.0 open-source announcement (2026-09-10) and related coverage 5. Industry daily roundup of world-model releases in the second week of September (2026-09-12)